“The slope of the power law for generalization error appears to be a property of the problem domain rather than the model architecture, according to Hestness et al. (2017).”